On 15/12/2021 13.51, Cornelia Huck wrote:
On Wed, Dec 15 2021, Thomas Huth <thuth@xxxxxxxxxx> wrote:
On 14/12/2021 22.55, Tony Krowiak wrote:
On 12/13/21 11:11, Cornelia Huck wrote:
One possibility is simply blocking autoload of the module in userspace by
default, and only allow it to be loaded automatically when e.g. qemu-kvm
is installed on the system. This is obviously something that needs to be
decided by the distros.
(kvm might actually be autoloaded already, so autoloading vfio-ap would
not really make it worse.)
Of the vfio_ccw module is automatically loaded, then the kvm
module will also get loaded. I startup up a RHEL8.3 system and
sure enough, the vfio_ccw module is loaded along with the
kvm, vfio and mdev modules. If this is true for all distros, then
it wouldn't make much difference if the vfio_ap module is
autoloaded too.
I think I don't mind too much if we auto-load vfio-ap or not - but I think
we should make it consistent with vfio-ccw. So either auto-load both modules
(if the corresponding devices are available), or remove the
MODULE_DEVICE_TABLE() entries from both modules?
I think we really need to take a step back and think about the purpose
of adding a MODULE_DEVICE_TABLE()... basically, it declares which types
of devices on a certain bus a driver supports, in a way that can be
consumed by userspace (after file2alias.c worked on it).
Userspace typically uses this to match devices it is notified about to
drivers that could possibly drive those devices. In general, the
assumption is that you will want to have the drivers for your devices
loaded. In some cases (drivers only used in special cases, like here),
it might be a better idea to autoload the drivers only under certain
circumstances (e.g. if you know you're going to run KVM guests).
My main point, however, is that we're talking about policy here: whether
a potentially useful driver should be loaded or not is a decision that
should be made by userspace. Not providing a MODULE_DEVICE_TABLE does
not look like the right solution, as it deprives userspace of the
information to autoload the driver, if it actually wants to do so.
Thanks, Cornelia, that's a very valid point, I didn't see it from this
perspective yet, but it sounds like the right way to handle this. So I think
my patch should be applied, and if we then agree that the module should not
be auto-loaded by default, it should be handled with the appropriate changes
in userspace instead.
Thomas
